Acceptable use

This policy applies to the agents you deploy, what you grant them, and what you do with the data they reach. It forms part of the terms.

It carries more weight on this product than on most, because what you are buying is the ability to give software the power to act. Most of what follows is not about protecting the platform — it is about the people your agents will contact.

Contact people who have not contacted you and have no relationship with you, or send unsolicited bulk messages. An agent makes this cheap, which is exactly why it is prohibited here.

Monitor your own team. Using agent data, run records or response timings to build a surveillance record of an individual is a breach, and the product deliberately provides no tooling for it.

Build a profile or character assessment of a person, or record judgements about individuals you would not be prepared to disclose to them.

Make a decision about a person's employment, credit, housing or access to a service without a human deciding it. An agent may gather and present; it may not decide.

Impersonate a person, or present an agent's output as having come from a named individual who did not review it.

Work around a refused capability — routing a refund through a custom tool you wrote, for example. The refusal is the policy, and circumventing it through your own service is a breach regardless of whether it succeeds.

Work around an approval requirement by enabling a broad category to avoid a narrow one.

Attempt to reach another organisation's data, another agent's secrets, or to probe the platform outside the responsible disclosure process.

A published agent must declare every capability it requires and must not request capabilities it does not use. A listing that understates what an agent asks for is a breach and will be removed.

You may not publish an agent that attempts to obtain capabilities through a route other than its declared request, or that behaves differently once installed than in review.

Grant the least an agent needs to finish its job. Broad grants made for convenience are the ordinary way this goes wrong, and a permission review exists to catch them.

Record only what you have a lawful basis to hold. People explaining a problem frequently volunteer far more than the situation requires, and an agent will retain whatever you allow it to.

Mark sensitive material as sensitive. The controls exist and using them is your responsibility.

You may not place instructions in messages, attachments, imported content or event payloads intended to change how an agent behaves — yours or anybody else's. That content is treated as data by design, and attempting it is a breach regardless of whether it works.

We would rather tell you about a problem than suspend an account, because suspending an agent platform can stop a business process mid-flight. Where we can, we will contact you first.

Where there is active harm to the people your agents are contacting, we may pause the agents responsible first and tell you immediately afterwards. We will pause the specific agents rather than your account wherever that is enough.